Sie sind auf Seite 1von 9

130 IEEE TRANSACTIONS ON AUTOMKTIC CONTROL, VOL. AC-26, NO. 1.

FEBRUARY 1981

Properties of Numerical Algorithms Related


to Computing Controllability
CHRIS C.PAIGE

Absfmct-% numerical propertics of some methods for computing and observable if and only if
conbollab~arellsedinanexpositorywaytomotmafeawtdernnder-
standing of numericaloomputstions. In particular,the numerical rank of a
rank(CT, A ~ c ~ , ,. (. A. = ) ~ - ’ c ‘ ) = ~ (3)
matrix, the numerical stability of algorithms, the sensitivity of prublem,
andthesdingofproblemsaredlscussed.Anmuericdiystable~
ia glven for computing cmdhbility, but it is pointed out W a meaoure where the superscript T denotes transpose.
of the distance of the given system fromthe nearest uncontrollablesystem R.2 [2], [3]: The system (1) is controllable if and only if
would be more usefi~I,and this appears to be an open computational
problem. rank(B,A-XiI)=n, i = l , - . -, n (4)

and observable if and only if


I. INTRODUCTION
rank(CT,AT-Aiz)=n, i = l , . - . n. (5) Y

T HERE are severalmathematicallyequivalent ap-


proaches to determining controllability or observabil- R3: Thesystem (1) iscontrollable and observable if
and only if there exists a matrix K such that the eigenval-
ity of time-invariant linear systems of the form
ues of
i =A x
nXn
+
Bu , y = C x
nxm qxn
(1) A BKC + (6)

-
where A has eigenvalues X,, * ,A,. Unfortunately, differ- are all different from each eigenvalue of A .
Theseresults can lead to severaldifferentcomputa-
ent approaches lead to computational methods which can
givemarkedlydifferentresults.Suchdiscrepancies are tional approaches, three of which are the following.
causedbytherounding errors that occurduringthe CI: Form the matrices in (2) and (3) and compute their
computation, and one aim of this paper is to give some ranks.
understanding of this behavior. Thus, although determin- C2: Computetheeigenvalues of A and compute the
ingcontrollability of (1)is a simpleenoughproblem ranks of the matrices in (4) and (5).
mathematically, it is not such a straightforward computa- C3 [4]: Compute the eigenvalues of A , compute a ran-
tionalproblem, and itbringsoutsomerathersubtle dom matrix K,compute the eigenvalues of A + BKC, and
numerical properties of algorithms. As a result, it is an compare these with thoseof A .
excellent problem for exhibiting several general aspects of The first five sections of this paper will point out some
numerical algorithms that may help the control engineer general properties of numerical algorithms as they relate
in the choice or design of reliable numerical algorithms totheabovecomputationalapproaches.Because of the
for more general problems. A second aim of this paper is similarity of the two problems, it will save space to con-
to examinethecontrollability of (1)moreclosely,since sider controllability alone, for example, taking C = I in (6).
some questions related to computing controllability of (1) The treatment for observability will be obvious. The re-
remainunanswered. A numericallystablealgorithm is mainder of the paper will describe a different computa-
given for computing controllability,but it is indicated why tional approach which is closely related to both R2 and
this is not really the full answer to the problem. R 3 , and wdl &scussscaling and makeacomment on
Three widely used mathematically equivalent results on deriving some measure of controllability. Any uncontrol-
controllability and observability of (1) are the following. lable system is in a certain sense arbitrarily close to some
RI [ 11: The system (1) is controllable if and only if controllable system,but a controllable system mayor may
not be close to some uncontrollable one. Thus, a measure
rank(B,AB,---,A”-*B)=n (4 ofhow far acontrollablesystem is from the nearest
uncontrollable system would be of far greater use than
Manuscript received April 11, 1979; revised June 13, 1980. This work just knowing that the system is controllable.
was supported by the Natural Sciences and Engineering Research Council The results R1, R 2 , and R3 are validfordiscrete
of Canada under Grant A8652. systems, and everything in this paper will also hold for
The author is with the School of Ccnputer Science, McGill Univer-
sity, Montreal, P.Q., Canada. such systems.

00 18-9286/8 1/0200-0 130$00.75 81981 IEEE


PAIGE: PROPERTIES OF NUMERICAL ALGORlTHMS 131

The paper is intended to be largely expository. It is not In a practical problem, the actual matrix B in (7) may
a surveypaper, and the references will besomewhat not be available; instead wewill often have a matrix
incomplete asa result of the author's limited knowledgeof and a measure of accuracy BTOL such that
the control literature. It is hoped that a combination of
numerical experience with some understanding of control IIB-B112fBTOL.IIBI12. (10)
systems w iallow the author to communicate some useful
l
A numerically stable algorithm [9] applied to 8 on a
numerical ideas to a fairly general audience in the area of
floating-point digital computer will give the singular val-
control systems. The paper is a rewrite and extension of
t26l.
ues 6, > --- > 6, not of & but of a nearby matrix B" where
it is known from a rounding error analysis that

Equations (2) and (4) for determining controllability are the scalar STOL being the product of the relative preci-
based on finding the rank of a matrix. This is a reasonably sion of the computer, the number of iterations required in
straightforward task computationally, and is most reliably the computation, and a low order polynomial in n. Note
done bycomputing the singularvaluedecomposition that
(SVD) of the matrix (see [7], [SI, and the implementation 11B-B"11, <(BTOL+STOL).IIBII,
in [9]). If we assume that n > m, the SVD of an n X m
matrix B is < TOL-6,
B=U , S , V ~ , TOL 2 (BTOL+STOL)/(l-STOL), (12)
u p , = v p ,= v,v: =I, , so we see from (8) that if
S,=diag(u,,..-,u,), U,>U,>-.. >u,>O (7)
where the superscriptH denotes complex conjugatetrans-
pose and can be replaced by T if B is real. Here the real then ur could not be zero, but is is possible that a,+,, -
,,a
scalars u1,...,,a are called the singular values of B, V,is a could be, and so we can say that B has effective rank r,
unitary matrix of right singular vectors,and the left singu- meaning we know that B has rank at least r , but we are
lar vectors are the columns of VI, which is only part of a unable to say anything more with t h i s precision of data
unitary matrix if m <n. and computation. If BTOL=O above, we sometimes say
An attractive property of the singular valuesis that they that B has numerical r a n k r. In general, B would have to
are not very sensitive to changes in the matrix. In fact, if be perturbed by a matrix with norm at least
6, > * > 6, arethe singularvalues of the perturbed
matrix B+6B, then (see, for example, [lo, p. 3211 or [ l l ,
.
6,.-TOL 6 , (14)
P- 251) to obtain a matrix of rank r - 1.
I u , ~ ~ , ~ < ~ ~ S Bi =
I Il ,~- -, - m ,
7 (8) Computing the SVD is fairly expensive, and another
approach to determining rank is to compute an orthogo-
nal decomposition

Thus, singular values are "well conditioned" with respect QHBP=( R So )


to perturbations in the matrix.
Ideally,ifu,>--- > u r > O a n d u r + , = . . . = u ~ = O , t h e n
where Q is unitary, R is upper triangular and nonsingular,
B has rank r, and from (8), for any j < r it will require a
,
change 11 6B 11 > 7 to produce a perturbed matrix B+6B
and P is a permutation matrix; see, for example, [ 11, ch.
31. In theory, the dimension of R gives the rank of B, but
of rank less than j . Infact, +
B SB 2 U, it can be difficult computationally to decide the effective
diag(u,,--. ,uj-,,O,. -.
,O)Vy is such a matrixwith rank of ( R , S ) , and the result does not give the distance to
11 SB 11 =q. Thus, the singular values of a matrix tell us the nearest matrix of lower rank. If economy of computa-
not only the rank, but also how far the matrix is from the tion is important, this approach can be tried, and if it does
nearest matrix of a given lower rank, and the SVD gives not give a clearenough result, the computation can be
us such a matrix. continued so that in (15), using a general unitary matrix
Any matrix is arbitrarily close to a matrix of full rank P,R is altered to R" and S becomes zero [ 11, ch. 31. The
as can be seenbyalteringevery zero a, in (7) toan singular values of I? are the nonzero singular values of B,
arbitrarily small value E> 0. So unless the elements of a and the smallest one can be estimated as in [12] or [13]. If
matrix are known to infinite accuracy, the exact rank of this result is still not satisfactory, the SVD of R" can be
the matrix will usually be unobtainable. But what is of computed.
interest is the closeness of the matrix to one of a certain The main purpose of this section wasto emphasize that
rank, and the singular values tellus just this. we cannot,in general, expect to find the true mathemati-
132 IEEE TRANSACTIONS ON AUTOMATIC CONTROL,VOL. AC-26, NO. 1. FEBRUARY 1981

cal rank of a given matrix, as this is meaningless when we an obviously controllable systemwas uncontrollable. In
are dealing with numbers of a limited precision. However, fact, a perturbation of about mustbemade to A to
it is straightforward to compute the effective rank of the obtain an uncontrollable system, a very large change in
given matrix, that is, the rank of a matrix which to within comparison with 10 - I 2 . Fortunately, we can do better by
the precision of data and computation could be the given using C2. It is fairly straightforward to show that in this
matrix.The S V D alsogivesareliable indicator of the case, the smallest singular value of ( B , A - A i l ) is greater
distance to the nearest matrix of lesser rank, and, in fact, than 2-(”+’) forevery i, and for n = 10, C2 would cer-
it ismoreuseful to know the computedsingularvalues tainly give the correct result on a computer with precision
than the “rank.” -’
10 or smaller.
Many of the ideas inthis section have already been Theweakness of C1is that the originalproblem is
presented in the control literature, for example, in [25]. transformed to the intermediate problem (2) which can be
Much of the early work on the subject appeared in the sensitive to far smaller perturbations than would affect
statistical and numerical literatures. either the original problem or the intermediate problem
(4), and it is widely recognized that C1 must be discarded
as a general computational approach. As is often the case,
111. SENSITIVITY
OF PROBLEMS
one good example is sufficient to condemn a poor algo-
Perturbation analysis is the study of the sensitivity of rithm. To pinpoint the difficulty, the example was chosen
the results of problems to changes in data. We saw in (8) so that the matrix in (2) could be accurately computed,
that small changes in 3 led to small changes in its singular showing that it was the very small singular values of this
values, and so the singular value problem is well condi- intermediate matrix that led to the trouble; the problem
tioned. On the other hand, if we consider the solution of would often be made worse by the rounding errors that
equations problem A z = d with n X n nonsingular A , and would normally occur in the computation of this matrix.
perturb d to d + 6 d , the bound we obtain for the size of the A more general lesson can be learned from this exam-
change 6z in the solution (see [ 10, 1941) p. is ple, and this is the danger of transforming problems by
matrix multiplication. If the transforming matrix is ill
I I ~ z 1 l 2 / I l ~ 1 1 2 ~ x X ( ~ ) I I ~ ~ 1 1 2 /where
I I d I I 2 X ( 4 iscalled
the condition number of A for solution of equations. Here conditioned for solution of equations [see (16)], the trans-
the multiplying factor formed problem will usually be more sensitive than the
original problem. This is what happened in C1. The well-
known approach [14] of avoiding forming CTC in solving

minimizeIly-CxII,
is the ratio of the largest to smallest singular values of A ,
and the solution of equations problem for A is ill condi- is an important instance of eliminating this danger, and
tioned if this is large. has led, for example, to the introduction of “square-root”
Perturbation analysisisapurely mathematical study, filtering; see [15]. The warning against multiplication by
and the sensitivity of a problemis independent of how general matrices applies equally well to multiplication by
that problem is solved.However, the numerical analyst inverses, and it is advisable to avoid forminginverses
mustalwayskeep the sensitivity of problemsclearly in whereverpossible, and to arrange computations so that
mind when designing algorithms, for if in the process of any unavoidable solution of equations occurs as late in
solution the originalproblem is transformed to amore the process as possible.
sensitive problem, the error in the computed solution will However, not all matrix
multiplications should be
usually reflect this and be larger than necessary. avoided.Forexample, if x ( A ) in (16) isclose to unity,
This is the main weakness of computational approach then there is neghgible loss in forming AB. Since x ( Q ) = 1
C1. To exhibit this, consider the clearly controllable sys- for any orthogonal matrix Q, transforming data with
tem orthogonal matrices is quite acceptable, and is often the
best way of solving many matrix problems numerically;
see [7]-[ll], [14]-[18].

Iv. SENSITIVITY OF EIGENVALUES

The approaches C2 and C3 depend on finding eigenval-


Here ( B , A B , . . . ,A“-’B) has ( i ,j ) element 2(i-1)(1-’) ues, and so it is important to consider the sensitivity of
whichcanbeformed and stored withfullaccuracy on eigenvalues of general square matrices. This subject is
most computers. Even so, with n = 10 this matrix has its treated in depth in [ 161. Unlike singular values, the eigen-
three smallest singular values 0.613X lo-’’, 0.364X lo-’, values of some matrices can be very ill conditioned, that
0.712x 10 -’,
and its numerical rank will be less than 10 is, very sensitive to small changes in the matrix, and this
using the S V D on a computer with relative precision no can be the case even when the eigenvalues are well s e p
smaller than 10 - 12. In this case, C1 would conclude that arated. Wilkinson [16, p. 901 gives the following illuminat-

~~
~
~
PAIGE: PROPERTIES OF NUMERICAL ALGORITHMS 133

ing example of a 20 X 20 matrix: TABLE I


RESULTS OF ~ O M P ~ ~ A ~ O NTESTS
A L . ON (23)

‘20 20
19 20
-.32985+j1.06242 ,99999 .002
18 20
.. ..
A

A= .92191fj3.13716 -8.95872+j 3.73260 ,004

. . 3.00339+j4.80414 -5.11682+1 9.54329 ,007


2 20 5.40114+j6.178€4 - .75203+j14.148167 ,012
1, 8.43769zj7.24713 5.77659zj15.58436 ,018
11.82747+j7.47463 11.42828+j14.28694 .026

---
with well-separated eigenvalues Ai = i, i = 1, ,20. If we 15.10917+j6.90721 13.30227cj12.90197 .032
perturb 2 by a very smallquantity E in the (20,l) position, 18.06@36+j5.66313 18.59961cj14.34739 ,040

we obtain eigenvalues, to first order in E , 20.49720+j3.81950 23.94877+jll.80677 .052

.Ob4 8.45907
22.06287+j1.38948
28.45618Tj
Ai(E)%Ai +CY,€.
32.68478

Thedistressing fact is that these ai are verylarge, for


example,
a20= -aIG 4 X IO7 compute an eigenvalue f i i of A +BK+ SA’:
al0= -al1 = 4 x 10l2 11 6A’IJ2 Q ETOL. 11 A + BK 11 2 . (22)

so the eigenvalues of 2 are extremely ill conditioned. There is no reason to expect any correlation between
Numerically stable algorithms for finding the eigenval- 6A and 6A’, and even if p i = A i , we could have pi very
ues of a given matrix A [17], [18] will give the eigenvalues different from p i and xi
very different from Ai, and in
of a nearby matrix A” [16] where, in a similar manner to different directions, so that f i i and xi
look like distinctly
(1 different eigenvalues.
As an example, computations were carried out on the
IIA-A”112QETOL.IIAI12. (192 uncontrollable systemwith
However, unlike the singular value case, if an eigenvalue
Ai of A is ill conditioned, then we haveseen that the
xi
corresponding eigenvalue of A” may be very different. where 2 is the matrix in (IS) and Q is a random orthog-
An immediate conclusion ofthe foregoing is that it is onal matrix found by transforming a square matrix whose
poor practice to attempt to determine if a square matrix is elements are uniformrandomnumbers on (- 1,l) to
singularbycomputing its eigenvalues; it is the singular upper triangular form using the QR decomposition (see,
values that are required. Note that if A has singular values for example, [lo], [ 113) and taking Q as the transforming
u1> - > a, and eigenvalues ordered so that 1 X I I > * 2 matrix. The computedeigenvalues xi
of A and f i i of
IA, 1, then it is straightforward to show that +
A BK were obtained using single precision (six signifi-
cant hexadecimaldigits) on the AMDAHL 470 V7 at
McGill University. The elements of K were random num-
but these inequalities can bevery loose. For instance, with bers from a uniform distribution on (- 1,l). The results
large a, the matrix are given in Table I where, for comparison with C2, the
values

has I J ~ - I C Y ~ - ~ and ul-IaI. Thus, we see that nonzero


eigenvalues may give no indication of how close a matrix
is to singular.
The possibility of ill conditioned eigenvalues explains werealsocomputed. This valuewas the same for both
why approach C3 will sometimesgivewronganswers. eigenvalues of a complex conjugate pair.
Even if A and A + BK do have an equal eigenvalue X i , it In theory, one eigenvalueof A should be the same as an
w l +
i be difficult to decide this computationally when Ai is eigenvalue of A BK, this one being unity. But as can be
ill conditioned. Instead of finding the eigenvalue Ai of A , seen from Table I, the computed eigenvalues
xi
we will compute an eigenvalue of A”=A +SA where
of A” are xi
almost unrelated to the true eigenvalues of A and the
II II2 Q ETOL. IIA /I 2, +
computed eigenvalues f i i of A BK. A user seeing only
(21)
xi
the and f i i would probably conclude from C3 that the
and instead of finding an eigenvalue p i of A +BK, we wil systemisclearly controllable. The algorithm C2 fared
134 IEEE TRANSACTIONS ON ALTOMATIC COhmOL, VOL. AC-26, NO. 1, FEBRUARY 1981

little better. In theory, one value of p should be zero, but depend on their sensitivity to such equivalent changes in
there is only the slightest hint that some modes could be data.
“less controllable” than others. C2 was also carried out
with the true eigenvalue unity of A , and this gave the
VI. TRANSFORMATION
TO AN EQUIVALENT
SYSTEM
computed value in (24):
The aim here will be to transform the original system
p(B,A - I ) = 5 X l O - ’ , (25)
(1) to an equivalentsystem for which controllability is
indicating that it is the eigenvalue computation that caused obvious. Careful use w libe made of orthogonal matrices
the failure of C2, just as for C3. so that the computedsystem will be equivalent to a
A good method that is not based on an eigensolution systemveryclosetothe original system in the sense of
will be described in Section VI. It can be shown that this (26), and the algorithm will be numerically stable in this
method is numerically stable, and so Section V will briefly ideal sense. The algorithm will be presented for the more
consider numerical stability as related to the problem of general model,
determining controllability. E i =Ax +Buy E nonsingular, (27)
since little w li be lost in simplicity.The mathematical
STABLE
V. NLJMERICALLY ALGORITHMS
formulation of a realistic system will sometimes take this
It was mentioned that a numerically stable algorithm form naturally (see[19,p. 1213), and itwill usuallybe
for finding theeigenvalues of a given matrix A can be preferable numerically to transform E , A , and B rather
relied on to give eigenvalues which are exact for a nearby than to workwiththe often morepoorly conditioned
matrix A”, with “nearby” defined in (19). To understand model of the form (1) with E - ’A and E -‘B; see Section
this further, we note that the most we can hope for when 111.
working with floating-point numbers on a finite precision Themodel (27) w l
li be transformed by orthogonal
computer is to obtain thecorrectanswer for a set of matrices P , Q, Z to give
numbers which is in some sense close to the given data;
for example, our data values are usuallyslightly altered E (Q~EZ)(Z=~)
just by storing them in the computer. In line with these
ideas, wewould liketo find a numericalalgorithm for = ( Q=AZ)( Z‘X) + (Q=BP)(P ~ U )
determining controllability which,when applied to the k &+&j
data A and B in (l), would always give the exact answer
for a nearby system: A
where for some integer k Q n and no = m
i=(A+6A)x+(B+6B)uY
116-4II2 GflOL.IIAII*,
GCTOL.IlBIl2Y
II6Bll2 (26)

the scalar CTOL being the product of the relative preci- Lo Ekk
sion of the computer, the number of transformations
required in the computation, and a low-order polynomial
in the dimension of the system. For a reliable algorithm, a
careful rounding error analysis would seek to prove sucha d = / A02 ‘ .
A=
result, and perhaps give an expression for CTOL which
mustbe independent of A and B apart from their di- Eii and Aii are ni X n i , i = 1; ,k , (30)
mensions.
There exist effective algorithms for which it cannot be A,,i-,isn,xni_,ofranknj, i=l;..,k-l. (31)
said that the computed results are exact for slightly per- Here 2 is block upper triangular and A” is block upper
turbed data; nevertheless, such results hold for most basic Hessenberg. Since E is nonsingular, the Eii are too. This
matrix computations. The majorwork of realizing the transformation is a direct computation rather thanan
importance of this approach in understanding the numeri- iterativeone,such as must be used in an eigenvalue
cal properties of algorithms and of developing the tools computation, and it stops when either
and techniques for rounding error analyses was carried
out by Wilkinson in [16] and elsewhere, and for many A k , k - , has rank n k , meaning that (27) is
algorithms, a feweasy observations and a reference to controllable, (32)
Wilkinson’s work are all that is required to prove numeri-
cal stability. Note that a result of the form (19) by itself or
tells us nothing about the accuracy of the computed Ak,k - =0, meaning that (27) is not controlla-
eigenvalues. The accuracy of thecomputedresults will ble. (33)
PAIGE: PROPERTIES OF NUMERICAL ALGORITHMS 135

T o justify this criterion, note that the criterion (4) for nonzeromatrices in the bottom block of rows on the
(1) is equivalent to right-hand side of (39).
Of course, the transformations are applied to the com-
rank ( B yA - A I ) = n for allh (34) plete matrices in (29), the final subscripts in (39) corre-
since if A Z A , , then A - A I has rank n. If we transform sponding to those in (29). The full k steps can be visual-
(28) to the form (11, of then (34) becomes rank ized by considering (29). In the first step, A,, is finalized
(E-'& k - ? l - A I ) = n for all A, which is equivalent to and the k matrix becomes upper triangular. In the second
step, A,,, and E , , are finalized and therest of the k
rank (B",A " - ~ k ) = n for all A , (35) matrix again becomes upper triangular, and so on. In the
kth, step Aif)k- is either zero or of full row rank, and the
so this is a criterion for (28) tobecontrollable. If the last two blocks of columns of and A" havebeenfinal-
matrices in (29) are substituted in (35) and (32) holds, ized. Thus,the kth steponly entails computations to
then (35) holdssince A,,,; ,A k , k - ] all havefullrow check if is
zero or has full row rank. Obviously,
rank and (28) is controllable. If (33) holds, then (35) will k < n , and this is a finite computation. In practice, this
not be true for A an eigenvalue of the generalized eigen- algorithm would be refined somewhat for efficiency.
problem If reliable orthogonal decompositions are used in this
AkkZ=AEkkZ, z#o (36) transformation, then a straightforward application of the
rounding error results of Wilkinson E161 shows that the
and the system is uncontrollable. The transformed system computed matrices in (29) are exact for a system which is
(28) is controllable if and only if the original systemis, so very close to (27):
now we just have to show how to obtainthe forms in (29),
(30),and (31).
Note that if E = I in (27), then taking Z = Q gives E"=I
, (30), and, in fact, for models of the form
in (28), (29)and
(l), we would use
Z=Q (37)
and save time bynot carrying out the computation QTEZ. Here CTOL is as describedfor (26). Since the controllabil-
The followingalgorithm description is easier to under- ity of the transformed system is obvious, it means that we
stand initially if this is assumed. will decide the controllability of a system very close to the
There are several ways to transform the system to the given system (27), and so this is, by definition, a numeri-
form (28)-(31). The algorithm here requiresk major steps, cally stable algorithm for deciding controllability.
the first being on Finding full row rank Ai,i-, requires a rank determina-
tion in (39), and as always, a tolerance must be used for
testing for numerically zero elements.If a computation of
the form (15) for this reduction does not give sufficient
confidence, then it should be augmented, or replaced, by
the singular value decomposition (SVD). In either case,
using a numerically stable SVD, the result (40)holds.
Finally, if the system does turn out to be uncontrolla-
ble, then the generalized eigenproblem (36) could be solved
to determine if the system is stabilizable.The uncontrolla-
ble subspace is,obvious from the form of (28) and (29).
This algorithm will give the correct answer for a nearby
system, but to compare its actual performance with C2
and C3,it was applied to the same ill conditioned prob-
(39) lem (23)using the same precisionof computation. Since in
(23) B is a column vector, the Ai,,- in (29) are scalars,
with A,, ,-, as in (31) and E$+') and being niX n , . and the present algorithm w i never require the S V D and
l
Orthogonal Q, and Pi are chosenpurelyto transform w libe much faster than C1,C 2 , or C3.Ideally,
to full row rank A i , , - ] , for example, as in (15). If
Af'i-, is already zero or has full row rank, the algorithm
stops with k = i. Otherwise, this ith step is completed by
choosing 2, to give upper triangular QTE$)Zi which can The computations gave A1,, = 4.35887;A , = 8.29969; 17
be partitioned as in (39). Since E$) is nonsingular, this last ,-, --
< IA,, I < 22, i = 3,4, ,19; and Azo,19 = 0 . ~ 7 ,
is a straightforward computation; see [lo], [l 11, or [14]. showing that the system is uncontrollable to this precision
The Qi and Zi are also applied to A;:) in (39). If the of computation. In light of the earlier computations with
algorithm did notstop this general step in repeated on the C2 and C3, this was a satisfyingly good result. To see if
136 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. AC-26,NO. 1, FEBRUARY 1981

this was just luck, the algorithm was applied to (23) with mations will not alter the sensitivity of a problem. On the
100 different random orthogonal matrices Q. In all cases, other hand, general coordinate transformations with Di in
similar good results were obtained. (41) being illconditioned for the solution of equations will
Matrix transformations like (39) form some of the tools usually not only alter the sensitivity of the problem, but
of the numericalanalyst's trade, and related algorithms introduce excessive rounding errors as well.
are used by Wilkinson [20], [21] and Van Dooren [22] in Settingup a mathematical model so that the mathe-
analyzing the generalized eigenvalue problem,and by Van matical problem accurately reflects the sensitivity of the
Dooren [23] in analyzinggeneral linear systems. An earlier physical problem may not be straightforward. The correct
work by Tse et al. [24] also suggested reducing the system choice of coordinate axes w i hopefully be clear from the
l
to the form of (29), but instead of using orthogonal physics of the problem, but the choice of scaling is often
transformation matrices [24] uses matrices which can be not so clear since physical quantities can be expressed in
very illconditioned for the solution of equations, and the different units. The scaling may depend on the particular
method appears to be numerically unstable. The effect on problem being solved for the model. For the controllabil-
numerical computations of general transformations and ity problem, Moore [5] suggests that if the state variables
scaling in particular will be examined further in the next are in close correspondencewith physical variables, then a
section. useful scaling would be "on a per unit basis so that if in
(1) it happens that JFx?dt is relativelysmall for some
input u, wemay conclude that the associated physical
VII. COORDINATE TRANSFORMATIONS
AND SCALING
variable isonlyslightlyexcitedby the input." Perhaps
If D l , D z ,and D3 are nonsingular matrices, such a scaling could be obtained a priori by deciding on
the smallest significant change for each element of x , y ,
and I( and scaling so these are all equal. If the decision
1,= D , A D ; ' x , + D , B D ~ ' u , , y l = D 3 C D ; ' x l (41)
can be made as to what each smallest sigmficant change
is, then this scaling is trivial to apply, but it is not at first
ismathematicallyequivalent to (l), as can be seen by glance clear that this necessarilyachievesMoore's aim.
taking x , = Dlx, u1 = D2u, and y 1= D 3 y . This is called However, balancing variables like this seemsreasonable in
coordinate transformation. If the Di matrices are also that asignrficantchange in aphysical variable would
positive definite and diagonal, it is usually referred to as correspond to a signrficant change in the model variable.
scaling. With exact scaling, the elements of each Di are In the next section, an apparently different scaling will
chosen to be integer powers of the floating-point com- be suggested. This w li depend on the model constants A ,
puter numberbase, and in this case, a transformation B, and C rather than the model variables x , y , and u, but
'
such as D,BD; will introduce no rounding errors where seems justified for the situation to be considered.
here we consider B to be already stored in the computer. As a result of this discussion, it can be seen that the
However, when a numerically stable algorithm is applied best a numerical analyst can do is to design algorithms
to the scaled data, we find that evenexactscalingcan which are numerically stable for the data to which they
have a sigmficant effect on the computed results. are applied. However,such algorithms can only be as-
There are differing opinions as to what an ideal scaling sured of giving physically meaningful results if the data to
should be.Some of the confusion can beresolvedby which they are applied have been sufficiently well scaled,
realizing that there are two fairlydistinct reasons for and the axes for the mathematical model have been cor-
scaling. In the first place, the engineerwouldlike to rectlychosen and so are physicallymeaningful for the
choose coordinate axes (transformations) and units (scal- problem being examined. This choice is best handled by
ing) so that the mathematical problem accurately reflects the engineer abstracting the mathematical model from the
the sensitivity of the physical problem. For example, if the physical system.
physicalsystemisclose to being uncontrollable, the
mathematical model should reflect this accurately. Later
the numerical analyst might want to scale to minimize the VIII. DISTANCE
FROM AN UNCONTROLLABLE
effect of rounding errors on the computed solution. The SYSTEM
first reason is by far the more important one, and if the
mathematical problem accurately reflects the sensitivity of So far we have been asking whether a system is control-
the physical problem, then the numerical analyst should lable or not, and this has led us to ask, for example, if (4)
be careful not to alter this sensitivity in any way so that he holds or if it does not. Such yes-no answers are suspect
or she can compute this sensitivity accurately. The beauty since it is well known that any uncontrollable system (1)
of orthogonal transformations is that they do not alter with rn > O is arbitrarily close to a controllable system,
norms, and effectively only rotate axes,each axis whereas if we have to carry out general computational
maintaining its exact relation to the others throughout. As transformations, we have seen in Section V that we can at
a result, it can usually be shown that orthogonal transfor- besttest the controllability of a nearbysystem. This
PAIGE: PROPERTIES OF NUMERICAL ALGORITIMS I 37

nearby system may well be controllable even if the origi- that it may prove useful as a preprocessor for a numeri-
nal system is not, and so it is hardy sufficient to just ask cally stable way of computing a measure p ( A , B ) .
if a system is controllable.
One is thus naturally led to seek some measure of how
far agivensystemisfrom an uncontrollable one. One IX. CONCLUSION
obvious algebraic approach for (1) is to take
The preceding contents havebeensomewhat destruc-
y(A, B ) minimurnll(6Ay6B)II such that the sys- tive, setting up computations for determining controllabil-
+ +
tem defined by ( A 6A, B SB) is un- ity, and then showing some of their drawbacks. On top of
controllable (42) this, the previous section emphasized the weakness of any
such approach which seeks only to determine if a system
as a measure of controllability. This measure with either is controllable or not. Nevertheless, the reader has been
SA or SB set to zero might also be of interest. The taken on a tour of some important numerical ideas. The
measure y(A, B ) will be invariant under orthogonal trans- paperconcludedbydrawing attention to abasic, but
formations, but could be altered by nonunitary transfor- apparently unresolved, problem in controllability- that of
mations and scaling. A numerically stable computation finding areliable and efficientnumerical algorithm for
for finding p ( A , B ) would ideally give the exact result for determining the distance of the closest unconrollable sys-
a nearby system. But it is clear that y ( A , B ) is perfectly tem from the given oqe.
conditioned with respect to changes in A and By and so From a tutorial viewpoint, it is useful to reemphasize
the computed y(A, B ) wouldbeveryclose to the true two clear distinctions that must be made in understanding
value. the numerical properties of algorithms. The first is the
Such an approach would have the following practical distinction between the sensitivity of a problem, and the
advantage. Suppose that the model matrices A and B are numerical stability of any algorithm applied to the prob-
known to differ from the true system matrices 2 and B lem. The sensitivity of a problem is a mathematical prop-
within erty of the problem which is independent of which algo-
rithm is used to solve the problem. On the other hand, if
l l A - 4 1 2 9ATOL.IIAI12,IIB-BIJ,<BTOL.IIBII,, an algorithm is numerically stable, it is ideally numerically
stable no matter what the sensitivity of the problem to
and the computed result y(A, B ) is exact for a system as which it is applied. Muchofthepower of Wilkinson’s
in (26); then if reverse error analysis of algorithms comesfromclearly
y ( A , B)>(ATOL+CTOL)JIAII, +(BTOL+CTOL)IIBII,, separating thesetwoaspects. A successfulreverse error
analysis of a numerically stable algorithm will give bounds
the designer would beconfident that the system is control- on equivalent perturbations in the initial data, and these
lable. If this is not the case, there is the possibility that the bounds will be independent of the sensitivity of the prob-
system could be uncontrollable. lem. If it is then necessary to see what effect such equiva-
The measure p ( A , B ) is obtained by allowing all possi- lent perturbations could have on the computed solution,
ble perturbations, but if the abovebounds on model the sensitivity of the problem must be considered.
uncertainties are dominated by the uncertainties in just a The other distinction is between the scaling of a prob-
fewelements, then the above criterion is liable to be a lem and the numerical stability of an algorithm for its
pessimistic one-large changes in these few elements may solution. There are two distinct problems here that have
not even alter the controllability of the system. For this to be solved in order to give useful results. First a well-
problem, a good scalingthen appears to be to scale so that scaled model has to be formulated, and then a numeri-
as far as possible the uncertainties in the elements of A cally stable algorithm has to be used to solve the problem.
and B are all of the same order of magnitude. This follows A numerically stable algorithm on a computer of a given
the recommendation in [9, p. 1.1 11 for scaling in solution precision wil give as good a result as is possible for the
of linear equations. Of course, accurately known elements data it is given; however, the results will only be physi-
such as 0 or 1 are ignored in such scaling. cally meaningful if the scaling chosen was itself meaning-
It is salutary to note thatof the methods for computing ful. Thus, the scaling and choice of axesis the crucial
controllability that havebeenmentionedhere, only the front end and should depend on the physical properties of
new one in Section VI gave a useful measure of y(A, B ) the system.The numerical algorithm comes later and
in its application to the uncontrollable system (23). For depends on the mathematical characteristics of the whole
controllable systems, no way of obtaining such a measure class of problems.
has been hinted at here. Related theoretical ideas are A connection between these two distinctions is that the
developed in [6],but the development of a practical com- choice of scaling can alter the condition of the mathemati-
putational algorithm still appears to be an open question. cal problem. There is little point in choosing a scaling on
One possible advantage of the reduction in Section VI is the basis of making the resulting mathematical problem
138 IEEE TRANSACTIONS ON ACTOMATTCCONSROL. VOL. AC-26,NO. I , FEBRUARY 1981

well conditioned; it is far more important that the scaling \/[I41 G. H. Golub, “Numerical methods for solving linear least squares
problems,” Numer. Math., vol. 7, pp. 206-216, 1965.
and choice of axes be such to ensure that the mathemati- P. G. Kamhki, “Square root filtering and smoothing for discrete
cal problem accurately reflectsthe sensitivity of the physi- processes,” Ph.D. dissertation, Dep.Elec. Eng., Stanford Univ.,
Stanford, CA,1971.
cal problem. J. H. Wilkinson, The Algebraic Eigenoalue Problem. London:
The original version of this paper [26] was written Oxford Univ. Press, 1965.
J. H. W b n and C. Reinsch, Handbook for Automatic Conputa-
without the benefit of computed results, and partly be- tion, Vol. II. Linear Algebra. Berlin:Springer-Verlag, 1971.
cause of this, the weakness of C2 as exhibited in Table I B. T.Smith et al., Matrix Eigeqvstem Rourines-EZSPACK Guide,
2nd ed., Lecrure Notes in Comput. Sci., vol. 6. New York:
was not detected. A referee thoughtfully pointed this out, Springer-Verlag, 1976.
thereby initiating the present computations. D. G . Luenberger, Introduction to Dynamic Systems. New York:
Wiley,1979.
J. H. Wilkinson, “Linear differential equations and Kronecker‘s
canonical form,” in Recent A h c e s in Numerical Anaiys~k,C. de
ACKNOWLEDGMENT Boor and G. H. Golub, Eds. New York: Academic, 1978.
- , “Kronecker’s canonical form and the QZ algorithm,” Nat.
Phys. Lab., Teddington,Middlesex, England,Rep. DNACS 10/78,
I am very grateful for the time and guidance given me Nov.1978.
by M. Denham of KingstonPolytechnic, and by D. P. Van Dooren, “The computation of Kronecker’s canonical form
of a singular pencil,” Linear Algebra and Irs Appl., vol.27,pp.
Mayne, P. Antsaklis, and R. Vintner at Imperial College 103-140.1979.
where this work was originallydone. A. Laub and the -, ‘”hegeneralized eigenstructure problem applications in
Linear system theory,” Ph.D. dmertation, Katholieke Univ. Leuven,
referees gave a great many thoughtful and helpful com- Leuven, Belgium, May 1979.
ments which were invaluable in the rewrite. G . Miminis E. C. Y. Tse, J. V. Medanic, and W. R Perkins,“Generalized
Hessenberg transformations for reduced-order modelling of large
and B. Bhattacharya carried out the computations men- scale systems,” Int. J . Contr., vol. 27, no. 4, pp. 493-512, 1978.
tioned here. V. C. Klema and A. J. Laub, ‘The singular value decomposition:
Its computation and someapplications,” ZEEE Trans. Automat.
Contr., vol. AC-25, Apr. 1980.
C. C. Paige, “Computing controllability and observability of time
REFERENCES invariant linear systems,” Imperial Coll., London, England, Com-
put. and Contr. Rep. 79/13, Mar. 1979.
R E. Kalman, “Onthe general theoryof control systems,”in Proc.
1st ZFAC Congr., vol. 1. London: Butterworth, 1960, pp. 481-491.
M. L. J. Hautus, “Controllability and observability conditions of
linear autonomous systems,” in Proc. Kon. Ned. Akad. Wetensh.
Ser. A , vol. 72, 1969, pp. 443-448.
H. H. Rosenbrock, State-Space ana‘ Multioariable Theov.
London: Nelson, 1970.
E. J. Davison, W. Gesing, and S. H. Wang, “An algorithmfor C M s C . Paige was born in Sydney, Australia, in
obtaining the minimal realizationof a linear time-invariant system 1939. He received the BSc. (mathematics and
and determining if a systemis stabilizabledetectable,”ZEEE Trans.
Automat. COW.,V O ~ .AC-23, pp. 1048-1054, DE. 1978. physics) degree in 1960 and the B.E. (electrical
B. C.Moore, “Computational problemswithmodalanalysis,” engineering) degree in 1962, both from Sydney
presented at the 16th Annu AUerton Cod. Commun., Contr., University, and an Academic Postgraduate Di-
Comput., Oct. 1978. ploma in numericalanalysis in 1965 and the
- “Siigular valueanalysis of linear systems,” in Proc. I978 Ph.D. degree in computer science in 1971 from
ZEEh Conf. Decision Contr., San Diego, CA, Jan. 1979, pp. 66-73. the University of London, London, England.
G. Golub and W. Kahan, “Calculating thesingularvalues and He worked for the Australian Department of
pseudo-inverse of a mat&” J. SIAM Numer. Anal., ser. B., vol. 2, Supplyfrom 1962 to 1964where he wasin-
pp. 205-224,1965. volved in the mathematical and computer mod-
G. H. Golub and C. Reinsch, “Singular value decomposition and
least squares solutions,”Numer. Math., vol. 14, pp. 403-420, 1970. eling of systems, and was a lecturer at London University Institute of
J. J. Dongarra, J. R Bunch, C.B. Moler, and G. W. Stewart, Computer Science from 1967 to 1972. Since 1972 he has been with the
U N P A C K User’s Guide. Philadelphia, PA: SIAM 1979. McGill University School of Computer Science, Montreal, P.Q.,Canada,
G. W. Stewart, Znrroduction to Matrix Computations. NewYork: wherehe isnow an AssociateProfessor. In 1978-1979 hespenta
Academic, 1973. sabbatic yearwiththe Department of Mathematics,ImperialCollege,
C. L. Lawson and R. J. Hanson, Soloing Least Squares Problems. London, and hasspentsummerswiththe Departments of Computer
Endewood Cliffs. NJ: Prentice-Hall, 1974. Science and Operations Research at Stanford University, Stanford, CA,
N.-hderson and.1. Karasalo, “On computing bounds for the least
singular value of a triangular matrix,” BIT, vol. 15, pp. 1-4, 1975. and with the Department of Scientific and Industrial Research, Welling-
A..K. Cline, C. B. Moler, G. W. Stewart, and J. H. Witkinson, “An ton, New Zealand. His main research interests are related to the design
estmate for the condition number of a matrix,” SIAM J . Numer. and analysis of effectivenumericalalgorithms and the application of
And., VOI. 16, p ~ 368-375,
. 1979. such algorithms in engineeringand statistics.

Das könnte Ihnen auch gefallen